The Video Doesn’t Lie. Until It Does.

Your phone lights up with a voice message from someone you know. It is their voice, with their pauses and their slightly annoying habit of saying your name twice. They need a favour. It is urgent. You would help, of course. Anyone would.

For most of human history, that would have been the end of the story. A voice was a person. A face on a screen was a person. The effort needed to fake either was so great that nobody bothered, except in films, where we paid to be fooled.

That arrangement has quietly expired.

In January 2024, a finance employee at the Hong Kong office of Arup, the British engineering firm behind some of the world’s best-known buildings, joined a video call. The chief financial officer was on it, along with several colleagues he recognised. They talked through a confidential deal and asked him to move money. He made fifteen transfers, about US$25 million in total.

Everyone else on the call had been generated. Every face, every voice. Hong Kong police later described it as a multi-person video conference in which the only real participant was the victim.

What stays with you about the Arup case is not the sum. It is how reasonable the employee was. He did not click a suspicious link or wire money to a stranger. He saw his boss, heard his boss, and did what his boss asked. He trusted the most reliable instrument humans have ever had: their own senses.

It helps to be precise about what a deepfake is, because the word has become a kind of fog. It is audio, video or an image made or altered by artificial intelligence so that it convincingly shows a real person saying or doing something they never said or did. A few years ago this took specialists, time and powerful computers. Now a passable voice clone can be built from a short sample of someone speaking. A face can be animated from a handful of photographs, most of which we have posted ourselves.

The obvious response is: fine, I’ll be more careful. I’ll look for the glitch — the odd blink, the sixth finger, the mouth that doesn’t quite match the words.

The research is not encouraging. In 2024, psychologists led by Alexander Diel reviewed 56 studies of how well people detect deepfakes, covering images, audio, video and text. Average accuracy was 55.5 percent. For practical purposes that is a coin toss with a slight tilt. Training and AI assistance pushed the figure to around 65 percent. That is better, but it is not a number you would bet your savings on, or your freedom.

And the glitches are disappearing. Each generation of tools removes the tells that the previous generation’s articles taught us to look for. Advice to “watch the eyes” has roughly the shelf life of a smartphone.

Here is where it gets stranger.

The damage deepfakes do is not limited to the fakes. In 2019, two American legal scholars, Robert Chesney and Danielle Citron, pointed out a second effect. They called it the liar’s dividend. The better everyone understands that video can be faked, the easier it becomes for anyone caught on a real video to say: that’s not me, that’s AI.

You have probably already seen a mild version of this. Under almost any striking clip online, someone now asks whether it is real. Often that is healthy scepticism. But scepticism that applies to everything stops being a tool for finding the truth and becomes a way of avoiding it.

The argument has already reached a courtroom. In 2023, in a wrongful-death lawsuit over a crash involving Tesla’s driver-assistance system, the company’s lawyers suggested that recordings of public statements by Elon Musk might not be reliable, since famous people are frequent targets of deepfakes. The judge, Evette Pennypacker, called the argument “deeply troubling”. If it worked, she observed in effect, anyone famous enough could disown anything they had ever said on camera.

It did not work that time. But notice the shape of it. A fake video fools you once. The idea of fake video lets anyone escape everything that was ever recorded.

That is the real inversion. For a century, the camera was how ordinary people held powerful people to account: the bystander’s footage, the leaked recording, the phone held up at the right moment. If every recording becomes deniable, the people who lose most are not the powerful. They are the ones whose only proof was a video.

Courts are catching up, unevenly. In September 2025, a judge in Alameda County, California, threw out an entire civil case, Mendones v. Cushman & Wakefield, after concluding that the plaintiffs had submitted AI-generated videos of a witness. Judge Victoria Kolakowski compared them with genuine footage of the same woman. The accent, the rhythm and the facial expressions were all wrong. The court, she wrote, had “zero tolerance” for this. It was a satisfying ending. It also depended on something most cases don’t have: real footage of the same person to compare against.

Legislators have mostly gone for labels. Since 2 August 2026, the European Union’s AI Act has required anyone who uses AI to create a deepfake to disclose that the content is artificial, and required AI companies to mark what their systems produce in a machine-readable way. It is a sensible rule. It is also a rule that fraudsters are unlikely to study closely. The Arup call did not come with a disclaimer.

The United States tried something more ambitious and discovered how hard it is. Federal judges proposed a new rule of evidence — Rule 707 — to hold machine-generated evidence to the same reliability standard as expert testimony. In June 2026, the committee overseeing the rules declined to move it forward and sent it back for more work. One of the problems was almost philosophical: the rule mainly covered evidence that someone admits was produced by a machine. A deepfake’s whole point is that nobody admits anything.

So if we can’t trust our eyes, and labels won’t save us, what’s left?

The answer turns out to be old-fashioned. Stop asking what a file shows. Start asking where it came from.

This is how courts have long treated physical evidence. A knife found at a scene is only as useful as the record of who picked it up, where it was kept and who touched it since. Digital files deserve the same suspicion and the same care. Is this the original, or a copy of a copy forwarded through three messaging apps? Who recorded it, on what device, and when? Does anything outside the file — a call log, a bank record, another person’s phone — confirm that the event it shows actually happened?

Technology is starting to help with the same idea. A standard called C2PA lets cameras and software attach “content credentials” to an image or video: a signed record of where it was captured and what has been done to it since. Think of it as a passport for a photo. It will not stop anyone from making a fake. It may make it much easier to show that something is real.

That shift, from spotting fakes to proving originals, sounds technical. It is really a change in who carries the burden. And that raises an awkward point. A company can afford forensic experts to authenticate its evidence. A tenant with a phone video of a threat usually cannot. If the answer to deepfakes is to make truth expensive to prove, we will have solved the problem for exactly the people who needed the least help.

In the meantime, the most effective defences are almost embarrassingly low-tech. Some families agree on a code word for emergency calls. Some companies now require any request to move money to be confirmed through a second, separate channel, however senior the face on the screen. Newsrooms are learning to publish where an image came from, not just the image.

None of this is glamorous. All of it rests on the same idea: that seeing something is no longer the same as knowing it.

The next time a video crosses your screen — a politician saying something outrageous, a celebrity in a place they shouldn’t be, a relative asking for help — the question may no longer be whether it looks real. Almost everything looks real now. The better question is whether anyone can tell you where it came from.

Sources: Hong Kong Police Force statements (Feb 2024) and Arup confirmation (May 2024), as reported by CNN · Diel et al., Computers in Human Behavior Reports (2024) · Chesney & Citron, California Law Review 107 (2019) · Huang v. Tesla, Santa Clara County Superior Court (April 2023), as reported · Mendones v. Cushman & Wakefield, Alameda County Superior Court, order of 9 Sept 2025 · Regulation (EU) 2024/1689, art. 50 · US Advisory Committee on Evidence Rules, proposed Rule 707 (2025–2026) · C2PA specification.